search

pacific-hake / hake-assessment

:zap: :fish: Build the assessment document using latex and knitr
MIT License
13 stars 6 forks source link

Trying to understand positive recruitments at age-0 in retrospective plot (Fig 55) #694

Closed iantaylor-NOAA closed 3 years ago

iantaylor-NOAA commented 4 years ago

My initial theory for why there are positive age-0 deviations is somehow related to the bias adjustment being applied to years with no information about recruitment due to the removal of the final years of data.

That is, the control file inputs

2018 # _last_yr_fullbias_adj_in_MPD
2019 # _first_recent_yr_nobias_adj_in_MPD

would need to be adjusted back by one year for each year of the retrospective. It might help to look at the "RecDev" group of r4ss plots for one of those retrospective models.

kellijohnson-NOAA commented 4 years ago

We adjust the main period, but we don't adjust the bias adjustment ramp inputs. Thanks for the input @iantaylor-NOAA

iantaylor-NOAA commented 4 years ago

Using a retro-6 model, I confirmed that shifting the inputs from

2018 # _last_yr_fullbias_adj_in_MPD
2019 # _first_recent_yr_nobias_adj_in_MPD

to

2012 # _last_yr_fullbias_adj_in_MPD
2013 # _first_recent_yr_nobias_adj_in_MPD

makes the recdevs in the recent years now match 0. Using MCMC for the retrospectives would also make the issue go away.

Digging deeper into why they were non-zero without that change, I realized that applying the bias adjustment to those years with no information about recruitment results in really low recruitment for the retro years (as shown in figure below) such that the realized catches can't be fully removed from the population. This is evident in the non-zero value for the "Catch" element in the table of likelihoods.

Changing the bias adjustment settings in future retrospectives should correct this although even without the downward adjustment to recent recruitments, there may still be cases where the catch can't be removed without positive recdevs.

compare9_recruits

aaronmberger-nwfsc commented 3 years ago

Good to see that our test evaluations on the bias adjustment settings during retrospective runs also confirms this behavior. We plan to move to mcmc for retros, but in the mean time need to remember to do this for MLE runs (see issue #728) where new lines of code will be added for retro runs.)

kellijohnson-NOAA commented 3 years ago

See #760 for an example of the squid plot and how this issue relates to MLE vs MCMC.